The Maximal Eigenvalue of O-l Matrices with Prescribed Number of Ones Shmuel Friedland* Department
of Mathematics
Cornell University Ithaca, New York 14853 and Califmia
Znstitute of Technology
Pasadenu, Califmia
9’1125
Dedicated to my teacher Beni Schwarz, who pioneered the subject 20 years ago.
Submittedby Richard A. Brualdi
ABSTRACT We determine the maximum spectral radius for O-l matrices with m2 + 1 ones for I = 2m,2m - 3 for aII m and for a fixed 1 and m > M( 1). Similar results are obtained for symmetric O-l matrix with zero diagonal. In all cases, equality is characterized.
1.
INTRODUCTION
Let G be an undirected graph on n vertices with 9 edges. Then G is represented by an n x n @l symmetric incidence matrix A = (aij);, having 29 ones and zero diagonal. We denote the set of such matrices by Sn,Bq.For A = (aij)E M,(C), let AP= (aIT)>. For A E Sn,2q, a$T) gives the number of distinct paths of length p connecting i to j. Assume that G is connected, i.e., A is irreducible. Suppose furthermore that A is primitive. Then a$T) = p(A)P[uiuj + 0(1)&P], for some 0 < E < 1, where Au = p(A)u, u = (U 1,. . . , u,,)~ > 0, IX:= luf = 1. Here, p(A) denotes the spectral radius of A. It is therefore important to find good upper bounds of p(A) in terms of 9. Let M n, k be the set of n x n @l matrices having k ones. Thus any A E M,, k
*On leave from the HebrewUniversity. LINEAR ALGEBRA
AND ITS APPLICATIONS 69X-69
0 Elsevier Science PublishingCo., Inc., 1985 52 VanderbiltAve., New York,NY 10017
(1985)
33 0024-3795/85/$3.30
SHMUEL FRIEDLAND
34
represents a directed graph with k edges (we allow the diagonal entries to be one, i.e., our graph can have l-length cycles). Then, for an irreducible and primitive A we have an analogous formula for aiy). In a recent paper Brualdi and Hoffman [l] considered the following maximal problems:
0.1) max p(A) = P(B*) A E K.2,
= o,,z4.
(1.2)
They found the values of P,,~ and u,,~, in the cases k=m2,
k=m2+l,
29 = m(m
- 1).
They also characterized the corresponding extremal matrices in these cases. Assume first that k = m2. Then any A, is permutationally similar to diag{ J,,O}. Here J, denotes a matrix whose all entries are equal to one. That is, the corresponding maximal graph is the complete directed graph on m vertices. For 29 = m(m - l), again the corresponding maximal graph is the complete undirected graph on m vertices. They also showed that for k = m2 + 1 the maximal matrices are obtained by inserting a useless additional 1 anywhere else. Those are only the maximal matrices except for k = 2,5. Let E, be the following (m + 1) X(m + 1) matrix:
E, =
p=
k=m2+l,
a:,= (oL~,...,qJ,
i=l,...,
a,=l,
9,
1 [ 2’
1
9=1-p,
(Y~=O, i=q+l,...,
(1.3)
m.
It is quite tempting to conjecture that E, is a maximal matrix. That is, pn.k
=
(1.4)
dEk)*
Indeed, we prove this equality for k = m2 +2m. k=m2+1,
Za2,
Also, we prove (1.4) for
maM(Z),
(1.5)
where M(2) is a big enough number. However for k = m2 + 2m - 3 the
MAXIMAL
EIGENVALUE
OF O-l MATRICES
35
maximal matrix is /
1 :
J,-i H m+1= 1 ,l
... ...
1 1
1’ :
i
i*
0 0
0 0,
(1.6)
In particular, (1.4) is false in general. We also show that H, is the exceptional matrix for k = 5. Let Hi + 1 be the matrix obtained from H, + 1 by replacing the ones on the diagonal with zeros. Then Hz + 1 is the maximal matrixfor2q=m2+m-2. These results suggest the following. CONJECTURE. Let
k=m2+1,
2q = m(m - 1)+2t,
l
lQt
Then there exist (m + l)X(m + 1) matrices A, and B, which satisfy (1.1) and (1.2) respectively. Moreover, if A, is symmetric, then B, is obtained from A, by replacing the diagonal elements in A, with 0. Besides the above mentioned cases, i.e. 1= 2m, 1 = 2m - 3, we also prove the conjecture in the case (1.5). We now outline the main ideas and results of our paper. The starting point of our paper and the paper of Brualdi and Hoffman is a fundamental theorem of B. Schwarz [6] which claims that A, can be chosen so that the elements of each row and column form a decreasing sequence. So A * is of the form where
B, = Jp.
In Section 2 we obtain upper estimates for p(F). Section 3 deals with the asymptotic expansion of p( Ek), k = m2 + I, m + co. In Section 4 we recall some basic inequalities connected with the rearrangement of vectors and prove a key lemma which is needed in the sequel. The computation of the values of pL, k in the cases we mentioned above are done in Section 5. Section 6 is devoted’to the characterization of the maximal matrices in M, k. The last section is devoted to the class Sn,s9. Here we combine our methods with the symmetric analog of Schwarz’s result given by Brualdi and Hoffman [l].
36 2.
SHMUEL FRIEDLAND PERTURBATION
RESULTS
As usual, let M,,(C) [M,,(R)] and M,,(C) [M,(R)] denote the set of m x n complex [real] matrices and the set of n x n complex [real] matrices respectively. We identify C”(R”) with M,,(C) [M,,(R)]. For A E M,,(C) let A’ and A* denote the transpose and the conjugate transpose of A respectively. For A E M,,(C) we denote by ]A], a(A), and p(A) the determinant, spectrum, and spectral radius of A respectively. We now give a sequence of lemmas which are needed in the sequel and may have interest of their own. LEMMA 1.
Let A, B E M,(C).
Assume that
(2.1)
p(A+B)>p(A). Then there exists z E C such that ~EO((~Z-A))~B), Moreover p(A + B) is the largest possible Proof.
Let IXI> p(A).
(2.2)
IzI > p(A). IzI
forwhich (2.2) holds.
Then
xZ-(A+B)=(XZ-A)[Z-(XZ-A)PIB].
(2.3)
As IXZ - Al # 0 for IX] > p(A), the determinant of XI -(A + B) vanishes iff (2.2) holds. Clearly p(A + B) is the maximum value of ]z] which satisfies (2.2). n LEMMA 2.
Let A, B E M,(C)
be of the
fm
A=(i2;), B=(; ;), A,,A;EM~,(C>, Then the rwn7zo
B,EMp(C),
P+VL
(2.4)
spectrum of A + B satisfies the equation lh2Z-AB,-A1A21=0.
(2.5)
MAXIMAL
OF Gl MATRICES
EIGENVALUE
37
Proof. Assume that h is a nonzero eigenvalue of A + B. Let 0 # (xt, yf)f be the corresponding eigenvector of A + B. That is, B,x + A,y = Ax, A,x = hy, X # 0. Clearly x # 0; otherwise the second equality would imply y = 0, which is impossible. So y = A,x/h, B,x + A,A,x/h = hx. That is,
( A21-
XB, - A,A,)x
= 0,
and (2.5) is established.
W
In fact, using a similar technique to that in [2, Vol. I, p. 461, one can deduce the identity
1AZ - (A + B) I= 1X21 - XB, - A,A,
I?I”~~~.
(2.6)
For B = (bij) we denote by B, the nonnegative matrix ( ]bjj]). Recall the Neumann expansion
(hi-A)-‘=
5
&,
lY>
p(A).
k=O
so Ak
[@Z-A)-‘]+< E 2=(IXIZ-A+)-1,
lXl>dA+).
k=OlhlkC1
(2.7)
LEMMA3. Let the assumptions of Lemma that A and B are nonnegative. Then
1 hold. Assume in addition
&)=&I-A)-‘B) is a strictly decreasing fin&ion unique solution of the equation
on (p(A),
p(r) = 1,
Proof.
XI). Moreover
p(A + B) is the
T ’ p(A).
(2.8)
Let
C(r)=(cij(r))=(rI-A)-‘B=
E
2,
k=O
*
r ’ p(A).
SHMUEL FRIEDLAND
38
Clearly, either cij( r) is identically zero or ci j( r) > 0 for all r > p(A). Let E = (eij), be the O-l matrix having zeros and ones at the places where cij(r) is zero and positive respectively. Assume that E is nilpotent. Then C(r) is nilpotent. So Ir~-(A+B)I=Irz-AI~z-(rl-A)-'BI=/rz-AI,
r > P(A).
Hence the characteristic polynomials of A and A + B are identical, which contradicts (2.1). According to the Frobenius form (e.g. Gantmacher [2]) there exists a permutation P such that /El
*
...
* \
0
E,
.--
* .
.
PEP’=
;,
0
..:
>
(2.9)
‘Eq
\ where each Ei is either irreducible or the 1 X 1 zero matrix. Thus, the nonzero spectrum of E comes from Eil,. . . , Ei; Let C(r) be permuted as in (2.9), and let C,(r) be the diagonal blocks i = 1,. . . , q. As each nonzero entry cij(r) is strictly decreasing we deduce that p(Cij(r)) is a strictly decreasing function on (p(A), co). Therefore
is a strictly decreasing function on (p(A), co). As p( A + B) is an eigenvalue of A f B, p(A + B) is a solution of (2.8). Since p(r) is strictly decreasing, W (2.8) has a unique solution for r > p(A). LEMMA 4. Let the assumption of Lemmu 1 hold. Assume furthermore that B is a rank one matrix. Then p( A + B) is the maximal 1z 1which satisfies the equation tr[(zZ-A)-‘B]
=l,
Zf in addition A and B are nonnegative, of the equation tr[(rZ - A) -‘B]
IzI ’ P(A).
(2.10)
then p(A + B) is the unique solution
= 1,
r ’ P(A).
(2.11)
39
MAXIMAL EIGENVALUE OF O-l MATRICES
Proof. As C(X) = (AZ - A)-‘B is also a rank one matrix, then the nonzero eigenvalue of C(h) is given by tr( C( h)). That is, p[(XZ-A)-‘R]
=]tr[(XI-A)-‘B]
1,
X4a(A).
Therefore, if A and B are nonnegative, C(r) is a nonnegative matrix for r > p(A) and p[(rZ - A)-‘B] = tr[(rZ - A)-‘B], T > p(A). Hence, according n to Lemma 3, p( A + B) is the unique solution of (2.11). LEMMA 5. Let the assumptions ofLemmas 2 and 4 hold. Then p(A + B) is the maximal IzI which satisfies the equation
tr[ z( z2Z - A,A,) In particular, solution of
= 1,
if A and B are nonnegative,
tr[ r(r2Z - A,A,)
Proof.
-1~,]
then p( A + B) is the unique
= 1,
-lo,]
(2.12)
M2 > P&42)
(2.13)
r2 ’ P(A,A,).
We first note that p”(A)
=
P@‘)
=
~(4-42)
=
pb424).
Next (X2~-hB,-A,A2)=(h2Z-A,A2)[Z-X(h2z-A,A,)P1B], 1x12 > p(A,A,).
(2.14)
According to Lemma 2 the eigenvalue X, [XI= p(A + B), satisfies (2.5). n Apply now the arguments of the previous lemmas. THEOREM 1.
Let
11.II be a submultiplicative nom on M,(C) ofLemma 2 hold. Then
(i.e.
IlABll G [[All ~~B~~). Let the assumptions
p(A+
and this inequality
B)<
.
11B,11+(11B,112+411A1A211)1’2 =11(A
is sharp.
2
>
B)
,
(2.15)
SHMUEL FRIEDLAND
40
Proof. inequality
It is well known that for any submultiplicative norm one has the
(2.16)
P(A) =GIIAII. See for example [3, p. 451. Hence P”(A) = &b42)
Q IIA1A211-
If p( A + B) G p(A), the inequality (2.15) is valid. Therefore we may assume that p( A + B) > p(A). The equality (2.14) and the obvious modification of Lemma 1 imply that P(A + B) is the maximal IX]> p(A) such that ~EO[~(~~Z-A,A,)~~B,].
(2.17)
The Neumann expansion, the triangle inequality, and the submultiplicativity of ]].]] yield
=
IhI2- “4A2”
’
Using the definition of 9 in (2.15), we deduce the inequality INIP1
II
l~12-I14A211
-Cl
for ]h] > n. In particular, (2.17) cannot hold for such a X. Hence we have (2.15). To show that (2.15) sharp we choose 11.1)to be the spectral norm ]]A]] = v(A) = ,o(A*A)“~=
(2.18)
P(AA*)“~
We then let B,=@u*,
/?>O,
u*u=l
MB,)
=
PI
(2.19)
41
MAXIMAL EIGENVALUE OF O-l MATRICES and A, = A;,
(2.20)
A,ATu = p( A,A;)u
so
[(r r2Z -
for
A,A,)
r=v(A,B).
According to (2.17) we have the equality sign in (2.15).
n
We improve the inequality (2.15) in the case that B, is a rank one matrix. For x E C”, denote by llxll the standard Z,-norm llxll = (C~==llxi12)‘/2.
THEOREM 2. Let the assumptions that B, is rank 1 matrix
B, = Buu*,
1 and 2 hold.
of Lemmas
@I>O,
u*u=1.
Assume
(2.21)
Let r = .$ be the unique positive solution of
IpI
f
k=O
Ju*b4,A2)k4
1
r2k+l
(2.22)
Then
Let R be the largest positive root of the equation R3-aR2-bR+c=O,
(2.23)
where
a=
IPI.
b = v(A2Al),
~=ISI[~~~2~,~-II~:~lIlI~2~ll].
(2.24)
SHMUEL FRIEDLAND
42
The-n
p(A+B)
Suppose,
in addition,
that A,A, A,A,
Then the above inequality equation (2.23) with
a = IPL
I
y*x = 1.
(2.25)
b = Ial,
c=(pa((l-lv*xlly*u().
(2.26)
is sharp.
The Neumann expansion yields
tr z z2Z-AA1A2)
[(
= my*,
holds, where R is the largest possible root of the
In both cases this inequality
Proof.
is a rank 1 matrix
m tr[ (AlA2)LB,] c z2k+l
%I=
k=O
=
f
Bv*b4~42)~u z2k+l
k=O
=sIPI
c O”
k=O
l~*b4,A2~k4 ,.2k+l
’
IZI
=
r.
Note that, since IPI > 0, the above series is a strictly decreasing function in r on the open interval (p(A), cc). So (2.22) has a unique solution 5. Since for r > 5, the values of the above series is less than 1, the condition (2.10) does not hold; hence p( A + Z?) Q 5. We next use the norm inequalities
MAXIMAL EIGENVALUE OF O-l MATRICES
43
Therefore
Itr [(z .z2Z-AA1A2) =-+IPI r
d r[r2-v(A2A,)]
= IPl[r2 - v(A,A,)]
+d
r[r2-r(AsAr)]
’ d =
I P III 4~
I( II A2u
II
Let R be the largest positive root of (2.23). So for r > R the right hand side of the above series will be less than 1. Therefore the condition (2.10) is not satisfied, i.e., P(A + B) < R. To see that this inequality is sharp choose 1/3I=/?>Oand A,u = A;v = z # 0,
Y(A,A,)z
= (A,A&
(2.27)
This will be the case if v=u#O,
A, = AT,
Assume next that A,A,
A,Afu = p( A,A;)u.
(2.28)
is rank 1 matrix of the form (2.25). Then
v*(A,A,)~u=~~(v*x)(~*u),
k>l.
Then the inequality p( A + B) < R is deduced as above. Clearly if
IPI>(Y>,O,
(v*x>(y*u) >,0,
(2.29)
we then deduce that p(A+B)=R.
3.
ASYMPTOTIC
Let Fmwl be
n
EXPANSIONS the following (m + l)X(m
+ 1) matrix:
(3.1)
44
SHMUEL FRIEDLAND
where J,
is the matrix whose all entries are 1 and og is defined in (1.3). Clearly F,, P, Q satisfies the assumptions of Theorem 2 with
B,=l,, A,A,
Al=+
= apa:,
A, = af,
(a rank one matrix).
Then (3.2)
Pm,p,g =P(FVw7) is the largest positive root of the equation (2.23) with a=m,
b = min(p,
q),
c=m[min(p,9)l-p9=min(p,q)[m-max(p,q)]. The equation (2.23) can be deduced directly. Indeed, the matrix F,,,,p, ~ has at most three distinct rows. So its rank is at most 3. Moreover, for 1~ p, 9 < m these three rows are linearly independent. So, in general, F,,,, p, 4 is a rank 3 matrix and the equation (2.23) is its characteristic equation IRI - F,,,, p, 91= 0 divided by Rme3. Since F,,,, p, 4 is irreducible and primitive, p,, p, 4 is a simple root of (2.23). To estimate p,, p, 4 we shall use the fact that pm,_ is the unique positive solution of (2.22). We shah employ the following obvious lemma.
LEMMA 6. Let f, be strictly a decreasing function Assume
on (ai, CQ), i = 1,2.
that a2 < r.
(3.3)
Suppose that fi’(ri)=l,
Ui < ri,
i = 1,2.
Then
Moreover if the inequalities inequalities in (3.4) are strict.
in (3.3) are strict for r = r,, i = 1,2, then the
MAXIMAL EIGENVALUE OF O-l MATRICES LEMMA 7.
Let m, p, q, s, t be positive
45
integers satisfying min(p, q) < min(s, t).
p+q=s+t,
ldp,q,s,tGm, Then
Pm,p,q
Proof.
<
(3.5)
Pm,s.t*
Put (3.6)
is the unique positive solution R > min(p, q) satisfying the equation sopm,p,q f,,,,,@)
(3.7)
= 1.
As pq < St,
min(p, q) < min(s, t),
we easily deduce f,,,,,(r)
< fm,,,tW
for
r2 > min(s, t);
then the inequality (3.5) follows from Lemma 6. LEMMA 8.
Let m, p, q be positive 2
integers satisfying lGp,qGm.
Then Pmpg
P,pg>rl=m+ , ,
Pq
m2-min(p,q)
’
Pq
rt-min(p,q)
*
(3.8)
46
SHMUEL FRIEDLAND
In pat-tic&r
(3.9) Proof.
We claim that
m
(3.10)
lGp,q
Indeed, since Fm,p,qhas at most (m + 1)’ - 1 ones, we have the right hand side inequality. Also, as Fm,p,gis irreducible for 1~ p, q, the spectral radius of Cl p lr is strictly greater than the spectral radius of its principal submatrix 1,. This proves the left hand side inequality of (3.10). So
f,,,,,,w
r[m2_zn(p
q)]
’
r>m.
As g(r2) = 1, Lemma 6 implies the first inequality of (3.8). In a similar way
f,,,,.,ww=y.+ pq r[$--min(p,q)] ’
r
and Lemma 6 implies the second inequality of (3.8). The expansion (3.9) follows from the inequality
n
4.
REARRANGEMENTS
We now recall some basic facts about the rearrangement of vectors. For U=(Ui,...,Un)fER” we denote by u_ =(u;,...,u;)* a vector Pu for some permutation matrix P such that {u,: } is a decreasing sequence. For
MAXIMAL EIGENVALUE
OF O-l MATRICES
47
u, 2, E R” we say that 27 majorizes u (u < u) if
iu;<$u,,
i=l
c
i=l
k=l,...,n-1,
i=l
c
i=l
See for example Marshall and Olkin [5] for a good reference on the subject. It is well known (e.g. [5]) that
u’u < l&u_,
(4.2)
and the equality sign holds iff there exist a permutation P such that PU=U-,
Pu=u_.
(4.3)
In fact, as the set of all u satisfying (4.1) for a fixed u is a convex set spanned by Pu, where P ranges over the set of permutation matrices, we easily deduce a generalization of (4.2).
Let a, /3, u, u E R” be given uectms. Assume that
LEMMAS.
and the equality
sign holds iff there exists a permutation Pa=u_,
matrix such that (4.6)
P/3=u_.
LEMMA 10. LetA\ and A, be k X p O-l matrices with 1, and 1, ones respectiuely.
Denote by e the vector whose all components
li=sip+ti,
ogsi,
o
are 1. Put
i = 1,2.
(4.7)
SHMUEL FRIEDLAND
48
Then etAlA2e Q w:w2,
The equality
sign holds iff
wi=
(,p _p,ti,o _ ,...)
QfA,R
(4.8)
(.
1 ,...)
_
= B,.
(4.9)
si rows equul et, the si + 1st TOWis of
Here Bi is the matrix in which the first the fom
Proof.
0)’i = 1,2.
forsomepermutation matrices P, Q, R
PA,Q = B;,
and the remaining
)...,
l,o )...)
0).
rows contain only zeros.
It is quite obvious that A\e < B,e,
A,e -C B,e.
Then the inequality (4.8) follows from (4.5). The equality (4.9) is implied by n (4.6). In [4] Katz studied a related maximum problem max e “A’e. * EM,,, He found the maximal value and characterized the maximal matrices when up to a permutation) and k = n2 - Z2,
k = m2 (the solution is diag{ J,,O} where l2 > n2/2.
LEMMA 11.
Let A E M,(C). p(A)
The equality
Then [tr(AA*)]1’2=
sign holds in all inequalities
[tr(A*A)]1’2.
iff aA is a hermitian
(4.10)
rank one
49
MAXIMAL EIGENVALUE OF O-l MATRICES matrix for som4? (aI = 1. In particular p(A)
A
The equulity sign holds in all the inequulities tionully similar to diag{ J,,,, O}.
l Mn.k.
(4.11)
iff k = m2 and A is permuta-
Proof. Clearly, we may assume that A # 0. As AA* and A*A are positive definite, all the eigenvalues of AA* and A*A are nonnegative. The definition (2.18) implies that
The equality holds if AA* and A*A have at most one nonzero eigenvalue. So the equality holds iff AA* and A*A are rank one matrices. The Cauchy-Binet formula implies that A is a rank one matrix. That is, A is of the form (2.21). Then
v(A) = lPlll4l Iloll~
P(A)= IPL
The Cauchy-Schwarz inequality yields that p(A) = v(A) iff o = yu, y > 0. So A = Z]p]yuu* for some (a]= 1. We next note that tr( AA’) = tr( A’A) = k,
A
E
Mn,k-
Combine this equality with (4.10) to deduce (4.11). Assume that the equality sign holds in all the inequalities in (4.11). So A = (ai j) is a symmetric rank one matrix. That is, aij = uiuj. As a,, = uz is either 0 or 1, then ui is either 0 n or 1. Thus PAP’ = diag{ J,,,,O} and k = m2. Thus we have shown that Pn.d
and the maximal matrix A, result is proven in [ 11.
=
m,
is permutationally similar to diag{ J,,O}.
(4.12) This
LEMMA 12. Let the assumptions of Lemma 2 hold. Assume furthermore that B, = 1, and A, and A, are O-l matrices with 1, and 1, zeros respectively.
SHMUEL FRIEDLANC
50
Let si and ti be defined as in (4.7). Denote by R the Zurgest positive (2.23) where
a=
b = ( ZJ2)1’2,
P,
c = p(z,z,)1’2
- ( w:wJ1’2(
root oj
w;w2)lP. (4.13)
p(A+B)
Proof
Lemma 11 yields v(A,A,)Q
v(A,)v(A,)b~~=(Z,z2)"2.
(4.14)
We now apply Theorem 2 to deduce Lemma 12. Indeed,
Ip = puu*,
u=e, 6
So for k > 1
where the last part of the above inequality follows from Lemma 10.
5.
THE NONSYMMETRIC
n
CASE
We now recall the Schwan result, which can be stated in this case as foIlows: Consider the maximal problem (1.1). Then there exists an THEOREM 3. extremal A, = (aTi); such that the entries of each row and column of A * fm a decreasing sequence.
51
MAXIMAL EIGENVALUE OF O-l MATRICES
We give a short outline of a proof of this theorem in the next section. Our results are based on the following corollary to Schwarz’s theorem. COROLLARY. Let A * = (aFj); be a solution of (1.1) such that each row and column formsa decreasing sequence. Let p be the largest i for which az=l. Thatis,
a*PP=1 ’
al;+ ljcp+ 1j = 0.
(5.1)
Then i, j=l
UTj = 1,
>.**9 PY
aFj=O,
i, j=p+l
,...,n.
(5.2)
Proof. Since columns h and p+ 1 of A, form decreasing sequences, the equality (5.1) implies
a?‘B
=1
’
i=l
,*..,P,
a;,+
1)=O,
i=p+l,...,
n.
As the rows of A, form decreasing sequences too, we easily deduce (5.2).
n
The Corollary explains why we studied the spectrum of the matrix A + B satisfying the assumptions of Lemma 2. THEOREM 4.
Let k=m’+l,
16162m.
Then p(F)<
m+&ZZi 2
(5.3)
forany n
x n, O-l matrix F having k ones. The eqdity sign holds iff I= 2m and F is permutation&y similar to the matrix diag{ E,,O}. (Note that E, in this case is symmetric.)
Proof.
According to the Corollary, we may assume that
(5.4)
52
SHMUEL FRIEDLAND
Here tr( A,A:)
= I,,
tr( A,A’,) = I,,
k = pL2+ I, + 1,.
(5.5)
Now Theorem 1 implies pi
P+[p2+4~(A1A2)]1’2
~ P+[P2+4(Z1Zz)1’2]1’2
2
2
~~+[82+2(11+z2)11’2~f(co 2 f(p)
=
P+
bk - ~~1~‘~ 2
(5.6)
*
Next we note that 2f’( /L)= 1 - p(2k - p2) - 1’2 > 0, That is, f(p)
OgpL
strictly increases in the interval [0,&l.
As P < m,
P(F)G f(m), which is the inequality (5.3). Assume that the equality sign holds in the above inequality. That is, p = m. Furthermore, in the series (2.22) we must have the equalities
where [[All = V(A). In particular e’A,A,e
= ~~(1~1,)~‘~ = m
MAXIMAL EIGENVALUE OF O-l MATRICES
53
so
I,=l,=s
< s2.
Hence s = m and the equality sign holds in the above inequality. Lemma 10 implies that A, = At, and A, has exactly one row of ones. So F is permutationahy similar to the matrix given by (1.3). In the next section we shah show that any extremal A, must be permutationally similar to a matrix F of the n form (5.4).
THEOREM 5.
Let m > 1.
k=m2+2m-3,
(5.7)
Then
P(F) G
m-l+(m2+6m-7)“2 2
(5.8)
for any n x n, O-l matrix F having k ones. For m > 2 the equality sign holds iff F is permutationally similar to diag(H,,,,O}, where H,+l is given by (1.6).
Proof. We first note that the right hand side of (5.8) is equal to f(m - 1). Assume that F is of the form (5.4). We now use the arguments given in the proof of Theorem 4. First, if p < m - 1, we have strict inequality in (5.8). Second, if p = m - 1 and the equality sign holds in (5.8) we deduce that Z,=Z,=2(m-l),
etAlAze = m [2(m -
l)] .
According to Lemma 10, we may assume, after a suitable permutation of rows and columns, that A, and Ai have the first two columns consisting entirely of ones. That is, F is diag{ H,, l,O}.
SHMUEL FRIEDLAND
54
In that case, F is a rank 2 matrix whose spectral radius P satisfies the quadratic equation p2- (m - 1)P - 2(7n - 1) = 0.
(5.9)
So we have the equality sign in (5.8). Assume next that F is of the form (5.4) with p= m. Lemma 12 implies thatP(F) is bounded from above by 0 < R which solves the equation ;+
(m - l)(m - 2) R{R2-
[(m-l)(m-2)]“s}
=l.
Clearly, if m = 2 then R = P = 2. Assume that m > 2. Lemma 6 yields that P(F) < R if R is the positive solution of
m+ cm-N-2) R
=I
R[R2-(m-l)]
*
We claim that R = p. Indeed, (5.9) is equivalent to P”-(m-l)=(m-l)(p+l). so m+ P
(m-l)(m-2) p[ps-(m-l)]
,m+ P =
(m-2) P(P+l)
m(p+l)+m-2 P(P + 1)
= ps-2(m-l)+P+2(m-l) P(P +I>
= (m - l)P + P +G
- 1)
P(P + 1) =
1
.
This proves the inequality (5.8). The equality case will be discussed in the n next section. THEOREM 6. Let k and m satisfy the assumptions of Theorem 4. Assume that F is an extremal matrix to the problem (1.1) of the form (5.4). Then
m-flxj.4
(5.10)
55
MAXIMAL EIGENVALUE OF O-l MATRICES unless 1= 1 and m = 1,2. In particular,
P(A)G m
for
and the equality sign holds if either A has J,,, as its principal k = 2 M 5 and A is permutationally similar respectively to
diag((ri),O) 0~ Proof.
(5.11)
A E M,,rn~+~~
submatrix,
or
diag{H,,O}.
Theorem 4 implies that it is enough to consider the case (5.12)
l
Let F be of the form (5.4). In the proof of Theorem 4 we showed that
where f(p)
is given in (5.6). A straightforward calculation shows that for
f(i.+m
p
Suppose first that 12 2. Then
Therefore we have the inequality (5.10). Assume now that 2 = 1. Clearly p(F) 2 m. Thus, if p(F) > m we must have the inequality 1_1>m - fi = m - 1. So p = m. But then, either A i or A, is a zero matrix and
P(F) = P(L) = m. This establishes the inequality (5.11). The equality case will be discussed in the next section. n THEOREM 7. Let 1 >, 2 be fixed. Then there exists M = M(1) such that for m >, M(1) any maximal solution A * to the problem (1.1) is permutation&y similar either to diag{ E,,O} or to diag{ E:,O}.
SHMUEL FRIEDLAND
56 Proof.
Choose m to satisfy the inequality (5.13)
m>,21+fi.
Let F be an extremal matrix of the form (5.4). By Theorem 6 the inequality (5.10) holds. Assume that A 1 and A, have 1, and 1, ones. So I, + I, = k - p2 = 2p( m-p)+(m-p)2+Z<2p(m-p)+2Z in view of the inequality (5.10). Put zi=sip+ti,
i = 1,2.
O
(5.14)
Then the inequalities (5.13) and (5.10) yield s1 +
s2
<
2(m
-
p).
(5.15)
Using Lemma 10, we get etA,A2e G (m - p)p2 + +T~T~, 72=(m-p)2+Z-71.
(5.16)
The equality sign holds for Ai as described in Lemma 10 with I, and I, of the form Zi = (m - p)p + TV, i = 1,2
or
Zi= (m - p)p + 73_i,
i = 1,2.
(5.17) Also
MAXIMAL
EIGENVALUE
57
OF O-l MATRICES
Next we estimate the series (2.22):
cm
k=O
e’( A,A,)%
e’(A,A&
pc+1
r
,
r3
II~:~IIII~2~llII~2~,II
+
r3
GE+ r
r2k+l
k=2
+b42)e
r
II~~~IIII~~~llll~~~~Ilk~l
+ g
r3( r2 - lb24
e’(A,A2)e
11)
+ [(TII-~)~~+~Z~](Z~Z~)~‘~
r3
rq r2-
(z1z2)1’2]
.
Then for r > m the above series are majorized by
g(r) = ; +
e”(A,A,)e lm2
+ [(wL-~)~~+~Z~](Z~Z~)~‘~ ?7nq?n-
(zlz2)1’2]
Thus, if p(F) > m, we deduce from Lemma 6
e’(A,A2)e + [Cm- ~11-1~ +2Z21 (hZ2)1’2 m2
dF)w+
let
m2[ m2-
p=m-s,
t5 18)
(Z1Z2)““]
1~scdi.A~ ( z1z2y2Q
9
m2 - ( Z1Z2)1’2>,m2-(m-s)s-Z, e’A,A,e
< s( m - s)~+ Z2,
we get
P(F)
s(m - s)~+ Z2 + [s(m
cm---+0
m2
s2 m
1 i m2 i ’
- s)~+~Z~] [(m m2[m2-(m-s)s-Z]
- s)s + Z]
*
SHMUEL FRIEDLAND
58
So for big enough m we get that p(F) < m, which contradicts our assumptions. Thus s = 0. In that case (5.18) yields
P(F)
Grn+---
e’A,A,e m2
In view of (5.16) P(F)<
wm +. 1
,m+ ww-
i m4 1 ’
m2
(5.19)
and the equality sign can hold only if A, and A’, have the same nonzero column, which have ri = [Z/2] ones in common. So F is permutationally similar either to diag{ E,,O} or to diag{ E:,O}. Finally, use (3.9) to deduce the equality sign in (5.19) for diag{ E,,O} and diag{ E:,O}. Thus we have proved the theorem for a matrix F of the form (5.4). The equality case for an arbitrary F is discussed in the next section. n
6.
THE EQUALITY
CASE
A matrix E is called reducible if E is permutationally similar to a matrix of the form (2.9) with o >, 2. The matrices Ei are called the components of E. Otherwise E is called irreducible. E is called completely reducible if the matrix given by (2.9) is block diagonal and each Ei is either an irreducible matrix or the 1 x 1 zero matrix. LEMMA13.
Let
m2+1
(6.1)
Then m=p,,,~+~
...
<1~.,(,+~)2=m+l.
(6.2)
Moreover for k 2 m2 + 2 any extremul A, is either irreducible or permutationully similar to a matrix diag( E,O) where E is an irreducible matrix. Proof.
We first note that n >, m + 1. Also the equalities rn=j.6
n,m=+l~
P,,(m+1)2
=
m
+
I
59
MAXIMAL EIGENVALUE OF O-l MATRICES were established before. Let k = m2 +2. Then
Assume that A, is reducible. So
P(A,) = P(E) for some component of A,. If E does not have m2 + 2 ones, then
&L) = P(E)G ~Ln,m2+1= m, which is impossible. So E has m2 + 2 ones, and hence A, is permutationally similar to diag{ E,O}. In case A, is irreducible, we let E = A,. Then E is an I X 1 matrix with 12 m + 1. Let G E M,,, and the nonzero entry of G be situated in a place where E has a zero entry. As E is irreducible, we have the inequality P n,mz+2
=
P(E)
<
P(E
+
G)
G ~n,,z+3
=
P(%).
See for example [2, Vol. II, Chapter 131. As before, we deduce that either B, is irreducible or B, is permutationally similar to diag{ E,O} whose E is irreducible. Continuing in the same manner, we prove the lemma. n We now recall the basic ingredient in the Schwarz theorem [6]. LEMMA 14. Let E be an 1 X 1 nonnegative rows e:, . . . , e:, such that Eu = p(E)u,
u > 0,
irreducible
matrix with the
(6.3)
u=u-.
DenotebyE_ themutrixwiththerows(e,)Y,...,(e,,)’, i.e., theithrowof E ~ is the i th row of E rearranged in decreasing order. Then
p(E) Q P(E- )a Zf E_ is also irreducible,
then the equality
(ef)u=(er)-u,
(6.4)
sign holds iff i=l,...,
n.
We give a short proof of this fact for the reader’s convenience.
(6.5)
SHMUEL FRIEDLAND
60 Proof.
As (et)u<(et)_u_
=(et)_u,
we deduce that E-u
>, p(E)u.
Since u > 0, we get the inequality (6.4) even if E _ is reducible. If E _ is irreducible and the equality sign holds in (6.4), then E_u=p(E)u. See for example [2, Vol. II, Chapter 131. So we have the equalities (6.5). Vice versa, if the equalities (6.5) hold, we deduce the above equality. Since E ~ is irreducible and u > 0, we deduce the equality p( E _ ) = p(E). n Assume that E_ is irreducible. Then v > 0.
(E-)‘o=P(E_)o,
(6.6)
It is quite easy to prove, using the above equality, that v = v ~. Put F=
[(E-j’]
_;
(6.7)
then according to Lemma 14 (6.3) Moreover, Schwarz proves that the elements of each row and column of F form a decreasing sequence. Proof of the equality case in Theorem 4.
Assume that I= 2m. Suppose
that
p(A,)
=
m +(m2 +4m)“’
Then according to Lemma 13, A,
2
.
is per-mutationally similar to diag{ G,O}
MAXIMAL
EIGENVALUE
61
OF O-l MATRICES
and G is irreducible. We may assume that G satisfies the assumptions of Lemma 14 (otherwise consider PGP’ for some permutation matrix P). Then G_ is also an extremal matrix. Therefore G_ is irreducible and we have the equality case in Lemma 14. Let E = (G_ )‘. Again E is irreducible and extremal, Put F = E ~. As before, F is extremal and irreducible. According to Theorem 4, F = E,, )’ > 0 be the eigenvector of E,. Clearly k=m2+2m. Let u=(ur,...,u,+i u1=
. .
.
=u,>u,+1,
hence the equalities (6.5) imply that E = E_ = E,. So G_ = E, and therefore w G = E,. Proof of the equality case in Theorem 5. Assume that m > 2. Then 1 = 2m - 3 > 1 and we repeat the arguments of the equality case in Theorem 4. For m = 2 we get k = 5 = 22 + 1 and the equality cases discussed below. n
Proof of the equality case in Theorem 6.
Assume that
Assume first that A, is reducible but not completely reducible. So
m = P(A*) = P(E), where E is some irreducible component of A *. Since E is irreducible, Lemma 11 implies that E = J,,,. In that case A * has j, as its principal submatrix. Assume next that A, is completely reducible and J, is not its component. The arguments above show that A, has only one nontrivial component E such that
m = P(E),
EEM,,&+1
and E is irreducible. Again we may assume that E satisfies the assumptions of Lemma 14. Consider E_ . Assume first that E_ is reducible but not completely reducible. So E contains I,,, as its principal submatrix. That is, E_ has at most m + 1 nonzero rows. Since E is irreducible, E is an (m + 1) x (m + 1) matrix. Since E_ has exactly m2 + 1 ones and E _ has J, as its principal submatrix, then either E _ has a zero row or E _ has m rows with m ones and one row i with a one in the first column. Since E is irreducible, E (and E ~ )
62
SHMUEL FRIEDLAND
cannot have a zero row. So we are left with the second possibility. Assume that m = 1. Then
and the only irreducible corresponding matrix is
Suppose m > 1. Then all rows of E_ except the ith row are of the form (l)...) 1,o ,...) 0). The assumption that J, is a principal submatrix of E_ implies that i = m + 1. From the proof of Lemma 14 it follows that E-u
2 mu,
u=u_
>o;
=%I,
u1 2 mu,,
therefore u1=
.**
I > 0.
But then we can never get the equality Eu=mu
for an irreducible E. We now assume that E_ is completely reducible. Then E- is permutationally similar to diag{ ./,, l}. So E _ has one row with one nonzero element. As m > 1, we have a contradiction as before. Finally we shah assume that E_ is also irreducible. Let F = [(E _ )‘I _ . Again it is enough to assume that F is irreducible. So F is of the form (5.4). Therefore p= m - 1 and p(F) = f( m - 1). From the proof of Theorem 5 it follows that F = H,, 1. But then m = 2. As before, we can show that E = F. n Proof of the equality case in Theorem 7.
Let m > M(l),
1 > 2. Assume
that
dA,) = P(G),
k=m2+1.
According to Lemma 13, A * is permutationally similar to diag{ E,O} where E E Mj,k is irreducible. We may assume that E satisfies the assump-
MAXIMAL
EIGENVALUE
63
OF o-1MATRICES
tions of Lemma 14. Then E_ is also extremal and is irreducible or completely reducible. In the second case E_ is per-mutationally similar to diag{ E,,O}. So E_ and E have a zero row. That is, E is reducible, and we contradicted our assumptions. So E_ is also irreducible. Thus F = [(E_ )‘I _ is irreducible and of the form (5.4). Therefore F is either E, or EL. In particular, E is an (m + l)X(m + 1) matrix. Let
u=(u1>...,u,+l)>
Fu = p(F)u,
u,>u,2
>U,.l>O.
.**
Since u is unique, according to the proof of Lemma 14 Gu =
P(G)= P(F),
p(G)u,
G
= (E_)‘.
Assume first that F = E,. Then
u1=
*-
.
=up>up+l=
.*.
=u,>u,+1r
p=
1 [2.1
Then the equalities (6.5) imply that F = G. The same arguments apply if F = EL. Thus E_ is either E, or EL, and as before we deduce that E = E_ . n
The inequality (5.11) and the equality case are proven in [ 11. However, as the methods in [l] differ from ours, we decided to give our proof for the sake of completeness.
7.
THE SYMMETRIC
CASE
Denote by S,,, the set of 61 n x n symmetric matrices with k = 2t ones and zero diagonal. Let S: k be the subset of S,, k consisting of those matrices A = (a i j) such that whenever i < j and a i j = 1, then a kl = 1 for all k < 1 with k
8 (Brualdi, Hoffman).
max A=Sn.lr
Then PB,P’
E S,f,
Let
P(A)=P(B*)=s,,.
forsomepfmnutation P.
(7.1)
SHMUEL FRIEDLAND
64
Assume that B * = (b;) E S,l:k. By considering the maximal p such that b&,+i) = 1, we deduce that B* is of the form
(7.2) Here by I, we mean the p X p identity matrix. As ~(J,$,)=p-1,
(7.3)
Theorem 1 and the inequality (4.11) imply
(P-l)+@k-P2+lY2
=g(p)
p(A)<
(7.4)
2
As g(x) is a strictly increasing function on [0, -1, the maximal value of the right hand side of (7.4) is achieved either for /.L = m - 1 or Z.L = m. Comparing these two values, we deduce THEOREM 9. Let k=m(m-1)+1,
Then forany
(7.5)
Ogl=2t<2m.
AESn,k m-l+[(m-1)2+2Z]1’2
p(A)g
(7.6)
2 For I= 0 the equality matrix
sign holds iff A is pennutationully
similar to the
diag{J,-I,,O}. REMARK. For 1 = 0, this result is due to Brualdi and Hoffman [l]. We now state a version of Theorem 2 which is needed here. THEOREM 10. Let A and B be real nonnegative (2.4). Assume furthermore that B, = puvt - yz,
U,V&O,
du=l,
matrices
p>y>o.
of the form
(7.7)
MAXIMAL EIGENVALUE OF CL1MATRICES
65
Then p(A + B) is the unique positive solution of
(7.8) Moreover,
if
A, = A;,
v=u,o, then the equality Proof.
A,A:u = wu,
(7.9)
sign holds in (2.15), where II.II is the spectral norm.
Let r = p(A + B). Lemma 2 and (7.7) imply that Ir(r + y)Z - r/3uvt- A,A,J = 0.
Note that rap(B,)=P-~-0,
r >, p(A).
Hence r satisfies the equation
As in the proof of Theorem 2, we deduce that r satisfies the equation (7.8). Assume next that (7.9) holds. As u > 0, it follows that w = p2( A). Also v( B,) = j3 - y. Summing the series in (7.8), we get the equation
pr
r(r+y)--o
=l*
So the equality sign holds in (2.15). THEOREM
11.
I
Let k = m(m - 1)+2m
- 2,
m 2 2.
(7.10)
Then for any A E S,,,
P(A)
G
m-2+(m2+4m-4)“2 2
(7.11)
SHMUEL FRIEDLAND
66
The equality sign holds iff A is permutationully similar to diag{ Hz+ 1,O}: where Hz, 1 is obtained j%nn H,, 1 (given by (1.6)) upon replacing its diagonal by the zero diagonal.
Proof.
As in the proof of Theorem
5, it is left to consider the case p = m.
AS
from Theorem solution
10 we obtain that p(A) < R, where R is the unique positive
of
m+ R+l
m-l [R(R+l)-(m-l)](R+l)
Note that the right hand side of (7.11)
(7.12)
=” satisfies the equation
(m - 1)’ = T(T + 1) - 2(m - 1).
(7.13)
so
m+ r+l
(m-1) [r(r+l)-(m-l)](r+l)
,Z+_ r-t1
=
m(r
(r:l)2 +l)+l
r(r+l)+r+l mr+m+l
=mr+2m-1 form>2.Sop(A)
2. Assume that A is of the form (7.2), follows that
11 A\e )I2 =2(m
p = m - 1, and p(A) = r. It then
- l)2.
Lemma 10 implies that A\ has precisely two rows of ones. That is, A is permutationally similar to diag{ Hz + Ir 0). Assume finally that m = 2. Then k = 4, and the only matrix in SC, is the matrix diag{ Hi,O}. n
OF O-l MATRICES
MAXIMAL EIGENVALUE
67
We state the symmetric analog of Theorem 7. Denote by Ei the matrix obtained from E, by replacing its diagonal with the zero diagonal. THEOREM 12. Let I= 2t >, 2 be fixed. Then there exists M( 1) such that fw m > M(1) any maximal solution B * to the problem (1.2) is permutation&y similar to diag{ Ei, O}.
Proof.
As p( B *) > m - 1, the inequality (7.4) yields m
-
p
-
+
<
(I + *y2.
Estimating from above the series (7.Q we deduce that the positive solution r of
J-+ r+l
e’A,Af,e (r+l)[r(r+l)-z,]
(7.14)
=l
majorizes p( B *). Here I, is the number of ones in A,. Recall that 21, = k - /.P + p = ( m - /.&)(m +/L - 1)+2t. Put p = m - s. Then 2Z,=s(2m-s-1)+2t
=2s(m-s)+s(s-1)+2t.
So, for a big enough m, Lemma 10 yields
. Let R be the positive solution of m-s R+l
s(m - s)2+ [t + s(s - 1)/212 + (R+l)[R(R+l)-t-s(m-s)-s(s-1)/2]
=”
(7*15)
Consider first the case s = 0. Then Theorem 10 yields that R = p(Ez). Moreover, Lemma 10 implies that p(B) = R iff B is per-mutationally similar to diag{ Ef, 0).
68
SHMUEL
FRIEDLAND
As in Section 3, we deduce that
R=
p(Ez) =
(7.16)
so
P(E:)[P(Ei)+ll>m(m-I)+$+0
5
.
i
i
We claim that
A+ R+l
(R+l)[R;a+l)-r] m-s
s(m - s)‘+
‘XT-i+
[t + s(s - 1)/2]’
(R+l)[R(R+l)-t-s(m-s)-s(s-1)/2]
for s 2 2. Indeed,
let x=R(R+l)-t.
Hence,
it is enough to show that
sx+ t2 x
>-
u
for
X-V
x=m(m-l)--t+$+O
f i
, 1
where
u=s(m-s)2+
[
t+p
1
S(S-1)
2
2
’
s(s - 1)
v=s(m-ss)+p
2
.
The above inequality holds for
x><= Expanding
sv+u--2+
[ (sv+u-~2)2+4t2sv 2s
l/2
1
the square root in a Maclaurin series, we obtain
5=
sv+u--2 S
(1
+o +.
(7.17)
MAXIMAL EIGENVALUE
69
OF O-l MATRICES
As sv+u--2 s =
(m - l)?n - (s - 1)m + O(l),
we have established the theorem for s >, 2. Assume that s = 1. If A, has one column of ones, then B, has .J, - I, as its principal submatrix and we are back in the case s = 0. Assume finally that each column of A, has at most m - 2 ones. As in Lemma 10, we deduce eA,Af,e<(m-2)2+(t+1)2
for m big enough. Using the above arguments, v=m-1,
s=l,
we get the equality (7.17) with
u = (m - 2)2+ (t + 1)“.
so (-=v+u-t2+0 c: m(m
i
$
1
=( m-1)+(m-2)2+0(1)=m(m-3)+0(1)
- 1)+0(l),
and the theorem is proved in this case too.
n
Z would like to thank Binyamin Schwarz fm reading the manuscript carefully and making many useful renuzrks, and Don Coppersmith forsupplying counterexamples to the maxima&y of E, which inspired Theorem 5. REFERENCES 1 2 3 4 5 6
R. Brualdi and A. Hoffman, On the spectral radius of (0,l) matrices, to appear. F. R. Gantmacher, The Theory ofMatrices, Vols. I, II, Chelsea, New York, 1959. A. S. Householder, The Theory of Matrices in Numerical Analysis, Blaisdell, New York, 1964. M. Katz, Rearrangements of (O-l) matrices, Israel J. Math. 95.S72 (1971). A. W. Marshall and I. Olkin, Inequalities: Theory of Majorization and Its Applications, Academic, New York, 1979. B. Schwarz, Rearrangement of square matrices with non-negative elements, Duke Math. J. 31:45-62 (1964). Received 11 April 1984; revised 26 April 1964